It is a simple recommendation system where user does not know which items to pick from a large set of items. Each item has keywords/tags based on which items can be clustered. Goal is to show user top k recommendations according to interests he has chosen. Here user 'interest' can be list of strings. Are there any standard algorithms for this 1 to top k recommendation problem? It will be great if there are any free/open source libraries available just in case to avoid reinventing the wheel ;)..
You can form segments of users based on the various attributes of the users using clustering methods.
If you have history information, you can use association rule Mining like Apriori Algorithm. Map the formed clusters with association rules.
You may use content based recommendation algorithms (IDF, VSM or similar) for the same.
You may use a similarity based approach for recommending top-k solutions:
1. Compute the similarity scores between existing keywords/tags and user selected tags.
2. These similarity scores can then be sorted in descending order to recommend top-k solutions.
I agree with Dhrubasish and Sahil, you need some content based approach and most of them use similarity measures to evaluate how near is an object to the user preferences.Then, a weighted sum can be used to give you a final score to the resource and you can make a ranking.
An ad-hoc approach can be seen in Casali et al. A Recommender System for Learning Objects Personalized Retrieval, ERSAT, 2012
Best!
You can use user-user collaborative filtering strategy. Based on the purchase history of the user on the e-shop, compute his similarity scores with other users. Identify his KNN and make predictions based on what those users bought that he does not have in his basket.
You can use semantic similarity measures (web-based are my personally favorite) to obtain a good ranking. Maybe there are not free libraries to do it (I am not aware) but their formula are very easy to build with any language or tool.
Thanks folks for really useful suggestions.
thanks Ana for suggesting an awesome resource (A Recommender System for Learning Objects Personalized Retrieval). Yes, my problem statement is somewhat similar to the LO recommendations. The system shall guide user through recommendations.
1) Right now there is 'tag' data and 'category' data associated with 'items' and user have 'interests' data, so these 'interests' shall be mapped to 'item' clusters so that system can suggest user the items in that cluster. If 'items' can be clustered very well then the common tags from items within the cluster can be shown to user in a list of interests. So user can choose his interest. This can be primary way to map 'interests' to 'item cluster'. But then again there will the challenge to recommend items for 'interests' which could not be mapped directly to 'item clusters'.
*sample data is attached with this message.
2) I agree similarity based approach will be useful. I am trying 'jaccord coefficient' as a distance metric. Please suggest if you feel any better method to achieve tag based clustering.
If you are to employ similarity based approach then I think that the order in which the Tag attributes are arranged will matter alot. Furthermore, it is not clear if the users' interests data match with the item category data or the tag data.
In my opinion, I suggest that you should have an ordered Tag data concatenated with category data to give the item data matrix:
ITEM = {Tag category}
=> ITEM data will now become an N X M data matrix where N is the number of items and M is the number of attributes.
The elements of the ITEM data should be make numerical to allow for arithmetic computation i.e enrique = 1; peaceful = 2 and so on.
To get the model of a user's preference; assuming that the user's interest value matches with the item category, we identify all items whose category correspond with the target user's interest and we get the the mean value of each column. Then we now have the user's model as {mean(col1) mean(col2) .. mean(colM)}. To make recommendation of a new item to the user, we compute the similarity of the new item data matrix; remember {Tag Category}, with the user model, if they close so the recommendation should be made otherwise not.
Or we can have an item matrix containing all possible tag values as columns, such that if a tag value is present in an item row we use '1' and '0' otherwise. We use the same approach as I had already mentioned above to model the user interest vector.
That is;
To get the model of a user's preference; assuming that the user's interest value matches with the item category, we identify all items whose category correspond with the target user's interest and we get the the mean value of each column. Then we now have the user's model as {mean(col1) mean(col2) .. mean(colM)}. To make recommendation of a new item to the user, we compute the similarity of the new item data matrix; remember {Tag Category}, with the user model, if they close so the recommendation should be made otherwise not.
Hi Pratik, I am a teacher using a method called Infopragmatics, which is based on a linguistic algorithm (its core is a linguistic storm = controlled vocabulary matrix), supported by an uncertainty theorem. This method offers an efficient solution, but not limited, to those Spanish speaking users who try to get the most useful information from academic
databases which contents is in English. I posted in here, Research Gate, two papers from International Conferences. One paper presented in Crete, and the other in Istanbul. Users are quite satisfied to what they got from their own designed Linguistic Storms. I am sure these two works can contribute to obtain PERTINENT information (user's interest), rather than Relevant one (the system's strings). Wish you luck!
Hai,Pratik
follow my research article A New Similarity Measure for User-based Collaborative Filtering in Recommender Systems which i feel will give some solution to your problem.
Good Luck
Article A New Similarity Measure for User-based Collaborative Filter...
You are right Li Li. I miss-conceptualized it. That was why I made my second suggesion. Thanks
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Graphs